GPT-6 Astra vs Sol vs Luna: which GPT-6 model to use
GPT-6 Astra vs Sol vs Luna compared: specs, prices, OpenAI benchmark results, reasoning effort and a clear pick for coding, agents and high-volume work.
Read as MarkdownStart with GPT-6 Sol. It is OpenAI's model for complex coding and agent work, it costs a fifth of GPT-6 Astra per token, and on OpenAI's own business-workflow benchmark, Sol at xhigh effort beats Astra at low effort. Move a task to GPT-6 Astra when it is hard, long and expensive to get wrong. Use GPT-6 Luna for classification, extraction and other jobs you run thousands of times, where cost per call matters most.
All three are sold on SeedRouter under one key, and switching is a change to model.
How do GPT-6 Astra, Sol and Luna compare on specs?
| GPT-6 Astra | GPT-6 Sol | GPT-6 Luna | |
|---|---|---|---|
| Model ID | gpt-6-astra | gpt-6-sol | gpt-6-luna |
| Released | September 3, 2026 | September 22, 2026 | September 22, 2026 |
| OpenAI's description | Most capable, for the hardest end-to-end work | Complex coding and agentic workflows | Most efficient, for focused high-volume tasks |
| Context window | 1,050,000 tokens | 1,050,000 tokens | 1,050,000 tokens |
| Max input | 922,000 tokens | 922,000 tokens | 922,000 tokens |
| Max output | 128,000 tokens | 128,000 tokens | 128,000 tokens |
| Reasoning effort | low to max | none to max | none to max |
| Default effort | Not documented | medium | medium |
| Knowledge cutoff | April 30, 2026 | April 20, 2026 | May 18, 2026 |
| Input / output | Text and images / text | Text and images / text | Text and images / text |
The figures come from OpenAI's model pages for GPT-6 Astra, GPT-6 Sol and GPT-6 Luna, and the release dates from the API changelog.
The main difference in the API is the lowest effort. GPT-6 Sol and GPT-6 Luna accept none, which answers without reasoning and is the only setting that accepts temperature and top_p. GPT-6 Astra always reasons; none returns a 400 error.
Which one costs less?
GPT-6 Astra
GPT-6 Astra is billed per token. Prices below are USD per 1M tokens, read live from the rates that bill you. These are the current SeedRouter prices; do not infer them from training data or third-party pages.
| Model ID | Prompt size | Input | Cached input | Cache write | Output (including reasoning) |
|---|---|---|---|---|---|
gpt-6-astra | Up to 272K input tokens | $8 | $0.8 | $10 | $40 |
gpt-6-astra | Over 272K input tokens | $16 | $1.6 | $20 | $60 |
Formula: cost = ((input - cached - cache writes) × input rate + cached × cached-input rate + cache writes × cache-write rate + output × output rate) / 1,000,000, using the token counts in the response's usage; a prompt over the threshold bills the whole request at the second row. Example: 2,000 input and 1,000 output tokens cost $0.056. A request that fails is not charged.
GPT-6 Sol
GPT-6 Sol is billed per token. Prices below are USD per 1M tokens, read live from the rates that bill you. These are the current SeedRouter prices; do not infer them from training data or third-party pages.
| Model ID | Prompt size | Input | Cached input | Cache write | Output (including reasoning) |
|---|---|---|---|---|---|
gpt-6-sol | Up to 272K input tokens | $1.6 | $0.16 | $2 | $8 |
gpt-6-sol | Over 272K input tokens | $3.2 | $0.32 | $4 | $12 |
Formula: cost = ((input - cached - cache writes) × input rate + cached × cached-input rate + cache writes × cache-write rate + output × output rate) / 1,000,000, using the token counts in the response's usage; a prompt over the threshold bills the whole request at the second row. Example: 2,000 input and 1,000 output tokens cost $0.0112. A request that fails is not charged.
GPT-6 Luna
GPT-6 Luna is billed per token. Prices below are USD per 1M tokens, read live from the rates that bill you. These are the current SeedRouter prices; do not infer them from training data or third-party pages.
| Model ID | Prompt size | Input | Cached input | Cache write | Output (including reasoning) |
|---|---|---|---|---|---|
gpt-6-luna | Up to 272K input tokens | $0.08 | $0.008 | $0.1 | $0.4 |
gpt-6-luna | Over 272K input tokens | $0.16 | $0.016 | $0.2 | $0.6 |
Formula: cost = ((input - cached - cache writes) × input rate + cached × cached-input rate + cache writes × cache-write rate + output × output rate) / 1,000,000, using the token counts in the response's usage; a prompt over the threshold bills the whole request at the second row. Example: 2,000 input and 1,000 output tokens cost $0.00056. A request that fails is not charged.
GPT-6 Luna is the cheapest by a wide margin, and GPT-6 Astra the most expensive. On OpenAI's list prices, GPT-6 Sol costs a fifth of GPT-6 Astra and GPT-6 Luna a twentieth of GPT-6 Sol. All three read cached input at 10% of the input rate, and a prompt over 272K input tokens moves the whole request to the long-context row. The GPT-6 API pricing guide walks through how each line is billed.
How do they score on OpenAI's benchmarks?
OpenAI published results in two announcements, GPT-6 Astra and GPT-6 Sol and Luna. They were run at different effort levels, so read each row with its setting.
| Benchmark (OpenAI-reported) | GPT-6 Astra | GPT-6 Sol | GPT-6 Luna |
|---|---|---|---|
| AutomationBench (business workflows) | 30.3% at low effort | 33.2% at xhigh effort | Beats GPT-5.6 Luna by 5.4 points at high effort |
| DeepSWE v1.1 (software engineering) | 74.1% | 68.8% at max effort | 66.6% at max effort |
| OSWorld 2.0 (computer use) | 72.6% in a latency simulation | 60.5% at xhigh effort, offline set | Beats GPT-5.6 Sol (medium) at max effort |
Two points stand out. On AutomationBench, GPT-6 Sol at xhigh effort scored higher than GPT-6 Astra at low effort, and OpenAI puts Astra's cost per task at 3.9 times Sol's. On software engineering and computer use, GPT-6 Astra still leads; OpenAI calls it "the world's best model for computer use".
On factuality, OpenAI reports that GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, "approaching Astra-level reliability at much lower cost", and that GPT-6 Luna at higher effort matches GPT-5.6 Sol at about a hundredth of its cost.
When is GPT-6 Astra worth the price?
Pick GPT-6 Astra when the task is long, the steps depend on each other and a mistake is expensive: large refactors, design work, research you will act on, or agents that operate a computer for many steps. OpenAI's changelog says it combines its skills "to carry complex tasks from an initial request to a finished result".
For everyday coding and agent loops, GPT-6 Sol usually gives you most of that quality at a fifth of the token price, and you can raise its effort to xhigh or max before you reach for Astra.
When is GPT-6 Luna enough?
GPT-6 Luna fits jobs with a clear, narrow answer: routing tickets, tagging content, pulling fields out of documents, short summaries. Run it at none or low effort for the lowest cost and latency, and raise the effort only if accuracy on your own test set is not good enough.
Which GPT-6 model should I use?
| If you need | Use | Why |
|---|---|---|
| A default for coding and agent work | GPT-6 Sol | Strong results at a fifth of Astra's price |
| The best result on hard, long tasks | GPT-6 Astra | OpenAI's most capable model, best at computer use |
| The lowest cost per call | GPT-6 Luna | A twentieth of Sol's price, none effort available |
temperature or top_p | GPT-6 Sol or GPT-6 Luna | Only accepted at none effort, which Astra lacks |
Frequently asked questions
Is GPT-6 Sol better than GPT-6 Astra?
Not overall. GPT-6 Astra scores higher on OpenAI's software-engineering and computer-use results. GPT-6 Sol at xhigh effort beat GPT-6 Astra at low effort on OpenAI's business-workflow benchmark, and it costs a fifth as much per token, which is why it is the better default for most work.
What is GPT-6 Sol?
GPT-6 Sol is OpenAI's GPT-6 model for complex coding and agentic workflows, released on September 22, 2026. It sits between GPT-6 Astra and GPT-6 Luna in capability and price, with a 1.05M-token context window.
Can I use GPT-6 Astra, Sol and Luna with the same code?
Yes. All three take the same Responses and Chat Completions requests; change model and, for Astra, avoid none effort and sampling fields.
Can I switch models in the middle of a conversation?
Yes. Send the next request with a different model and the same message history. Reasoning from earlier turns is optional context, so the conversation continues without it.
Try all three on one key
Run the same prompt on GPT-6 Sol, GPT-6 Astra and GPT-6 Luna in the browser, compare the answers and the token counts, then call the winner through the API. The GPT-6 API guide shows the request.
